Is AI using more or less energy over time?

In August 2025, Google published a study on the energy of its Gemini assistant, accompanied by a blog post: in 12 months, the energy consumed by the median text prompt to the app was divided by 33. In June 2026, the same company published its annual environmental report: its electricity rose from 31.7 TWh in 2024 to 43.6 TWh in 2025, 37 % more. The 2 figures come from Google, and they describe 2 different quantities: the energy of a unit of service, and total consumption.

Google is, to our knowledge, the only provider to publish both 2 over neighbouring years, along with the volume of text processed. This is why this analysis relies on its figures. Microsoft, Meta and the International Energy Agency publish the same rise in the total; none publishes the energy of a query.

Google's study, a preprint by authors employed by the company, not peer reviewed, measures 0.24 Wh for the median text prompt to the Gemini app in May 2025. The boundary is that of the data centre: the AI chips, the rest of the server, the machines kept in reserve and the cooling. With the chips alone, Google would have obtained 0.10 Wh; the method used gives a value 2.4 times higher. The internet network, the user's device and the training of the models remain outside the count.

The study attributes the reduction to software: a factor of 23 comes from the models, a factor of 1.4 from better use of the machines. Our analysis of AI decarbonisation details these levers. 2 caveats limit the scope of the figure. The value is a median: the study states that small sets of long queries, or queries served by little-used models, "consume a disproportionate amount of energy", and the average is not published. And the comparison runs from May 2024 to May 2025: we found no energy-per-prompt measurement published since. The 2026 report gives a 78% reduction in the serving cost of Gemini in 2025, which is a cost and not an energy figure.

In May 2024, Google was processing 9,700 billion tokens per month across its products and APIs, a token being the unit of text that models handle. In May 2025, more than 480,000 billion, 50 times more. In May 2026, more than 3.2 quadrillion, a further 7 times more. These figures are announcements by its chief executive, Sundar Pichai, with no published method.

The 2 factors cannot be divided by one another. The first applies to a median prompt to an application; the second covers all of the company's products and its APIs, and the number of tokens per prompt may have changed. They show that volume has grown much faster than the measured efficiency. The only total published, electricity, confirms this.

Heavy uses are growing, and they weigh more per query. On a benchmark of open models, Hugging Face measures that reasoning models consume on average 30 times more energy than models without reasoning, and 150 to 700 times more for the same model tested with and then without reasoning. Anthropic counts, in its own data, about 4 times more tokens for an agent than for a conversation, and about 15 times more when several agents work together. Google's median is not sensitive to this type of query; the mean, which is not published, is.

The 2026 environmental report states that in 2025 the company experienced "the largest load growth in its history": 37% more electricity, after 27% in 2024. Its data centres rose from 17.4 TWh in 2021 to 42.4 TWh in 2025, 2.4 times more. The report links this rise to all of its products, AI included, and states that the pace was "tempered" by its efficiency gains on AI, without saying by how much, or what share of the total is AI.

The movement is not limited to Google. Microsoft's electricity use rose by 24% over its financial year ended June 2025, and Meta's by 20% in 2024. Worldwide, the International Energy Agency counts 485 TWh for all data centres in 2025, compared with 415 in 2024, and projects 945 in 2030 in its base case scenario; the projections chart for 2030 compares this scenario with the others, and our analysis of the electricity shortfall relates it to France's demand.

Electricity

Electricity use of Google and Microsoft

Electricity used by the company, in TWh a year · 2020 to 2025, published years

Source: Google, 2026 report; Microsoft, 2026 data sheet · projetcelsius.com/en/ai · CC BY 4.0

PNGCSV

The chart counts all of Google's electricity, offices included: 43.6 TWh and + 37% in 2025; its data centres alone, 42.4 TWh.

Google's emissions table gives 2 values for its electricity in 2025. Calculated with the carbon content of the contracts and certificates that the company buys, they fall by 3%, to 2.82 million tCO2e; this is the figure that the text of the report puts forward. Calculated with the average carbon content of the grids where its servers run, they rise from 11.07 to 15.15 million tCO2e, 37% more (our calculation). The difference between the 2 accounts, and what an electricity contract does or does not change, is explained in our analysis of AI decarbonisation.

In 1865, the economist William Stanley Jevons argued that more economical use of coal raises its consumption: if a blast furnace burns less of it per tonne of pig iron, the price of iron falls, demand increases, and the number of blast furnaces ends up more than offsetting the saving made by each. In Scotland, he noted, the coal burned per tonne of iron had fallen to less than a third, and total consumption had increased tenfold. At Google, the 2 movements go together; no published figure says what share of the rise in volume comes from the fall in cost.

For AI as a whole, the question remains debated. In the journal Joule, Microsoft researchers estimate in 2026 that progress in models, software and hardware could divide energy by 8 to 20, without saying, in the published abstract, from what starting point, and help keep demand at a manageable level; the same article notes that long reasoning and agents multiply the energy of a query by more than 10. Sasha Luccioni, Emma Strubell and Kate Crawford, in 2025, write that rebound effects "undermine the assumption that technical efficiency alone guarantees a net reduction in environmental harm". Our view on surface AI and deep AI draws the consequences for uses without an interface.

For us, at Projet Celsius, the energy of a prompt says nothing on its own about emissions, which depend on the total and on the carbon content of the electricity. A provider would make it possible to follow the real trend by publishing each year the average energy of a query, not only its median, and the electricity attributed to AI; neither Google, nor Microsoft, nor Meta publishes the second.